[m-rev.] for post-commit review: fix mantis bug 480 (mostly)

[Zoltan Somogyi]

On Tue, 6 Aug 2019 11:29:19 +1000, Peter Wang <novalazy at gmail.com> wrote:
> That looks okay, thanks.

I followed both your suggestions.

> Out of interest, am I right that since there is no obligation for a
> hypothetical Mercury compiler to perform CSE before switch detection,
> this disjunction is not necessarily a switch according to the definition
> in the reference manual:
> 
>     (
> 	Struct = struct(Foo1, _Bar),
> 	Foo1 = yes(_),
> 	...
>     ;
> 	Struct = struct(Foo2, _Bar),
> 	Foo2 = no,
> 	...
>     )

The language lawyer was Fergus, not me. Given that there is only one
Mercury implementation, I don't see much point in expending effort
in drawing an imaginary line in the sand between "what the language requires"
and what the compiler does. Even if a second Mercury implementation
*does* arise, *they* will decide what subset of Mercury they implement,
not us :-)

BTW, the cse pass is done *after* the switch detection pass. Its just that
if cse ever does change the code of a procedure, it repeats switch detection
on that procedure, then invokes itself again, until it gets to a fixpoint.
This is because changes by cse can find new opportunities for switch
detection, and vice versa. So cse is always sandwiched between
two invocations of switch detection.

Zoltan.